Fixes for the McBackend adapter

[michaelosthege]

What is this PR about?
This fixes two bugs that were reported by @thelogicalgrammar in pymc-devs/mcbackend#93.

The first was a leftover of a previous refactoring of sampler stats, and the way how the "tune" stat is handled in PyMC.

For the record: Our current approach of passing information about tuning/not-tuning via sample stats is brittle. It's a per-sampler thing, but in practice it's actually a per-iteration thing from the PyMC, ArviZ and McBackend perspective. IMO we should consider taking it out of the step method attributes and stats, and instead pass tune as a parameter to the astep method.

However, this is out of scope for this PR, so I aligned with the current design and added "tune" to two step methods that didn't have it, much like Ricardo did for Slice in fee9a02.

The second bug was just a small dtype thing for pickled stat objects.
I tested it locally with a ClickHouse server:

def test_issue_93_b():
    seconds = np.linspace(0, 5)
    observations = np.random.normal(0.5 + np.random.uniform(size=3)[:, None] * seconds[None, :])
    with pm.Model(
        coords={
            "condition": ["A", "B", "C"],
        }
    ) as pmodel:
        x = pm.ConstantData("seconds", seconds, dims="time")
        a = pm.Normal("scalar")
        b = pm.Uniform("vector", dims="condition")
        pm.Deterministic("matrix", a + b[:, None] * x[None, :], dims=("condition", "time"))
        obs = pm.MutableData("obs", observations, dims=("condition", "time"))
        pm.Normal("L", pmodel["matrix"], observed=obs, dims=("condition", "time"))

    ch_client = clickhouse_driver.Client("localhost")
    backend = ClickHouseBackend(ch_client)
    with pmodel:
        pm.sample(trace=backend, tune=5, draws=7, discard_tuned_samples=False)
    assert idata.warmup_posterior.sizes["draw"] == 5
    assert idata.posterior.sizes["draw"] == 7
    pass

Checklist

• Explain important implementation details 👆
• Make sure that the pre-commit linting/style checks pass.
• Link relevant issues (preferably in nice commit messages)
• Are the changes covered by tests and docstrings?
• Fill out the short summary sections 👇

Major / Breaking Changes

• None

New features

• None

Bugfixes

• Two bugs in McBackend adapter were fixed

Documentation

• None

Maintenance

• All of PyMC's step methods now emit a "tune" stat

---

📚 Documentation preview 📚: https://pymc--6835.org.readthedocs.build/en/6835/

[codecov]

Codecov Report

Merging #6835 (33364d5) into main (82c6318) will decrease coverage by 2.55%.
The diff coverage is 92.59%.

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #6835      +/-   ##
==========================================
- Coverage   92.04%   89.49%   -2.55%     
==========================================
  Files          95       95              
  Lines       16298    16322      +24     
==========================================
- Hits        15001    14607     -394     
- Misses       1297     1715     +418     

Impacted Files	Coverage Δ
pymc/step_methods/compound.py	97.17% <75.00%> (-0.46%)	⬇️
pymc/backends/mcbackend.py	100.00% <100.00%> (ø)
pymc/step_methods/metropolis.py	86.92% <100.00%> (+0.33%)	⬆️

... and 9 files with indirect coverage changes

[ricardoV94]

Not sure what's going on with the "object" vs "str" dtype. Otherwise LGTM

[michaelosthege]

Not sure what's going on with the "object" vs "str" dtype. Otherwise LGTM

thanks for the review!

To clarify this part: Sampler warnings are SamplerWarning objects, but as such they can't be stored in all backends. Therefore, the ChainRecordAdapter pickles them to a string, which can be stored in HDF5/ClickHouse/whatever.

That's why they must be announced as str-typed when setting up the backend